Esdebug5

[zasdfgbnm]

No description provided.

[github-actions]

Review updated until commit 918e1b5

Description

• Simplify MatmulOp output handling by removing redundant striding logic

• Refactor resetAllocationDomainAndContiguity to use tensor directly

• Update argument preparation to support expression evaluation outputs

• Introduce inferAllocationSizesAndStrides without validation for flexibility

---

Changes walkthrough 📝

	Relevant files
Bug fix	nodes.cpp Simplify MatmulOp output handling                                                csrc/ir/nodes.cpp Remove manual stride handling in MatmulOp::evaluate Directly return matmul_out without strided copy +1/-11
Enhancement	fusion_cache_utils.cpp Refactor allocation domain reset with tensor input              csrc/runtime/fusion_cache_utils.cpp Update resetAllocationDomainAndContiguity to take at::Tensor instead of sizes/strides Use inferAllocationSizesAndStrides to get sizes and strides Simplify contiguity update logic with direct indexing +15/-75  fusion_kernel_runtime.cpp Support expression evaluation in kernel preparation            csrc/runtime/fusion_kernel_runtime.cpp Add expression evaluation path for ExprEval scheduler Bind inputs and evaluate outputs using ExpressionEvaluator Pass is_expr_eval flag to updateWithSegmentOutputs +25/-7    tensor_metadata.cpp Split inference and validation of allocation sizes              csrc/tensor_metadata.cpp Rename inferAndValidateAllocationSizesAndStrides to inferAllocationSizesAndStrides Move validation to separate function call Add new inferAndValidate wrapper that calls infer function +12/-1
Documentation	tensor_metadata.h Declare allocation sizes inference functions                          csrc/tensor_metadata.h Add new inferAllocationSizesAndStrides function declaration Document purpose of non-validating inference function Keep inferAndValidateAllocationSizesAndStrides for compatibility +9/-0

---

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🧪 No relevant tests

⚡ Recommended focus areas for review Output Striding The PR removes logic that ensures the output tensor has correct strides by constructing a strided tensor when the inferred output is not contiguous. This may lead to incorrect memory layout assumptions in downstream operations. return {matmul_out}; Contiguity Update The refactored resetAllocationDomainAndContiguity function no longer handles mismatched allocation domains or updates contiguity in all cases, potentially leading to incorrect contiguity metadata on TensorView. const at::Tensor& tensor) { const auto [sizes, strides] = inferAllocationSizesAndStrides(tensor, tv, ExpressionEvaluator()); auto contiguity_without_reduction = computeContiguity(sizes, strides); std::vector<std::optional<bool>> contiguity; int64_t index = 0; for (auto id : tv->getMaybeAllocationDomain()) { if (id->isReduction()) { contiguity.push_back(std::nullopt); } else if (!id->isBroadcast() && !contiguity_without_reduction[index].has_value()) { contiguity.push_back(true); index++; } else { contiguity.push_back(contiguity_without_reduction[index++]); } } tv->setContiguity(contiguity); } ExprEval Output Handling When using ExprEval, output sizes and strides are computed via evaluation, but the updated contiguity logic may not be consistently applied due to changes in how resetAllocationDomainAndContiguity is called. if (is_expr_eval) { // For expr evaluated fusion, the striding rules follow that of ATen. ExpressionEvaluator eval_fusion; for (auto [i, v] : enumerate(group_to_run->inputs())) { auto tensor_pv = args_manager.checkTensorMap(v); eval_fusion.bind(fusion_to_run->inputs()[i], tensor_pv); } for (auto v : fusion_to_run->outputs()) { auto result = eval_fusion.evaluate(v); group_runtime_outputs.push(result); }

[zasdfgbnm]

!test